llama.cpp

Fix issue #4791 alloc causes compute_size to be calculated incorrectly in train-text-from-scratch, end result core dump

#5033

Merged